Writing

Agent theory · November 7, 2026

Expectation and Misconception in AI Tutoring

Bloom's 2 Sigma Problem defines the achievement gap between conventional instruction and one-to-one tutoring. Graesser et al.'s expectation-and-misconception framework shows how conversational AI agents can approximate that tutoring through targeted dialogue.

This essay examines how expectation-and-misconception tailoring, a conversational pedagogy framework, allows AI agents to approximate one-to-one tutoring. It is written for teachers and researchers designing or evaluating AI tools that assist students in classroom settings.

In 1984, Benjamin Bloom posed a challenge that has shaped educational research for four decades. He observed that students who received one-to-one tutoring combined with mastery learning performed two standard deviations better than students taught through conventional group instruction. This gap, known as the 2 Sigma Problem, established a clear theoretical justification for any technology attempting to scale individualized support. Bloom’s foundational paper asked whether methods of group instruction could ever be found that approach the effectiveness of personal tutoring. Today, AI agents operating in classrooms represent the most direct attempt to answer that question. But deploying a language model in a classroom does not automatically replicate the conditions Bloom described. The mechanism matters. To understand how an AI agent might actually close the 2 Sigma gap, educators need to look beyond the mere presence of a chatbot and examine the specific conversational structures it uses to guide student thinking.

A student working with a computer in a classroom

The Architecture of Conversational Pedagogy

Arthur Graesser, Natalie Person, and Jenny Harter detail how intelligent tutoring systems can use conversation-based pedagogy to approximate the interactive feedback loops central to Bloom’s mastery model. Their work focuses on a specific pedagogical strategy called expectation-and-misconception tailoring. When a human tutor asks a student a question, the tutor does not simply wait for any response. The tutor holds a mental list of expected concepts—specific points, principles, or facts that a complete answer should contain. Simultaneously, the experienced tutor anticipates common misconceptions, errors, or incomplete ideas that students frequently produce when grappling with the material.

Graesser et al. explain that effective conversational tutors structure their dialogue around these two lists. If a student provides an answer that hits an expected point, the tutor acknowledges it and prompts for the next missing expectation. If the student offers a known misconception, the tutor intervenes directly, addressing the error before moving forward. This back-and-forth is not random conversation. It is a highly structured diagnostic loop. The AI agent functions less like a search engine retrieving information and more like a clinician checking symptoms against a known profile.

For teachers coordinating work in a classroom, this distinction is practical. An AI agent built on expectation-and-misconception tailoring does not merely generate text; it evaluates student input against a pedagogical map. When a teacher assigns an AI assistant to help a small group work through a physics problem, the agent needs to know what correct reasoning looks like for that specific problem, and where students typically stumble. Without that architecture, the AI risks validating incorrect logic or providing explanations that bypass the student's actual confusion.

A teacher observing a student using an AI tool

Closing the Loop Toward Mastery

The connection between conversational pedagogy and Bloom’s 2 Sigma Problem lies in the concept of mastery learning. Bloom demonstrated that the massive achievement gains of tutored students were not solely due to the presence of another person in the room. The gains resulted from the combination of individualized attention and a requirement that students master each step before advancing. Conventional classrooms often move forward on a fixed schedule, leaving students behind when they fail to grasp a concept. One-to-one tutoring pauses, diagnoses the misunderstanding, and repairs it.

Graesser et al. show how expectation-and-misconception tailoring operationalizes this pause-and-repair cycle in a digital environment. By systematically checking off expectations and dismantling misconceptions through dialogue, the AI agent enforces a form of mastery. A student cannot simply skim past a difficult concept because the conversational structure requires them to address it. The agent asks follow-up questions, offers hints, or presents contradictions until the student articulates the expected idea correctly.

This process matters significantly when an AI agent assists a teacher rather than replacing one. In a typical classroom, a single teacher cannot maintain a continuous diagnostic dialogue with thirty students simultaneously. The teacher must rely on whole-group instruction, occasional questioning, and summative assessments to gauge understanding. An AI agent using the framework described by Graesser et al. can handle the granular, turn-by-turn conversational work with individual students while the teacher manages the broader instructional environment. The agent acts as a scalable mechanism for the interactive feedback loops that Bloom identified as essential.

Practical Constraints in the Classroom

Implementing expectation-and-misconception tailoring requires careful design. The AI agent must be equipped with accurate expectations and realistic misconceptions for the specific tasks students are performing. If the domain knowledge is shallow, the conversational loop breaks down. The agent might accept a flawed answer because it lacks the depth to recognize the error, or it might repeatedly prompt a student for an expectation that is poorly defined.

Teachers and researchers must also consider the pacing of these interactions. Bloom’s original findings relied on human tutors who naturally adjusted their tone, patience, and timing based on the student's emotional state. While Graesser et al. outline the cognitive and pedagogical mechanics of conversational tutoring, the affective dimension remains a challenge for AI. An agent that rigidly cycles through a checklist of expectations without recognizing student frustration may undermine the very motivation required for mastery learning.

Ultimately, the 2 Sigma Problem is not just a historical benchmark; it is a design specification. Bloom told us what level of achievement is possible when instruction is individualized and mastery-oriented. Graesser et al. provided a concrete conversational framework for building systems that attempt to reach that level. For educators integrating AI into their practice, the value of an intelligent tutoring system does not come from its ability to generate fluent sentences. It comes from its capacity to hold a structured dialogue—one that knows what to expect, knows what to correct, and refuses to move on until the student has mastered the step at hand. That is how an AI agent begins to function not as a novelty, but as a genuine approximation of the one-to-one tutoring Bloom envisioned.

Sources